FlowMap-OPD: Rollout--Kernel Separation for On-Policy Distillation of Few-Step Flow-Map Generators
Read the original on Hugging Face Trending Papers →The Flow has not summarised this story yet — read it at Hugging Face Trending Papers.
The Flow has not summarised this story yet — read it at Hugging Face Trending Papers.
arXiv:2609.37851v1 Announce Type: cross Abstract: Few-step flow-map generators, including MeanFlow and consistency models, enable efficient sampling through long-range transport, yet their on-policy...
The paper introduces Trajectory-Regularized Distribution Distillation (TRD), a method that merges distribution-level and trajectory-level supervision for one-step generative models. TRD uses a momentum-mixture policy derived from a single student forward pass, eliminating the need for an extra network to estimate the student’s score. Experiments on SANA-1.6B and Stable Diffusion 3.5 demonstrate that TRD surpasses existing consistency and distribution matching baselines in generation quality and training efficiency, though it slightly reduces sample diversity.
The paper introduces Velocity Adversarial Distillation (VAD), a method that uses a learned critic to compare velocity fields of student and teacher diffusion models, turning disagreements into learning signals. VAD unifies several existing distillation techniques and leads to VA‑TDM, which applies trajectory distribution matching and transition‑aware online velocity learning to supervise partial denoising along the student’s sampling path. Experiments demonstrate that VA‑TDM achieves state‑of‑the‑art performance for four‑step text‑to‑image and text‑to‑video generation, outperforming prior few‑step methods and approaching the quality of many‑step baselines.
Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamic...
arXiv:2607. 24522v1 Announce Type: new Abstract: While on-policy distillation (OPD) effectively addresses sparse rewards and exposure bias in large language model post-training, its extension to flow models remains underexplored.
The paper investigates how the critic in Distribution Matching Distillation (DMD) should be designed to provide a reliable training signal. Experiments on ImageNet with seven SiT model sizes show that keeping a small critic while increasing teacher size can degrade student quality and cause training collapse; scaling the critic together with the teacher (joint scaling) prevents this collapse and lets students benefit from larger teachers. Additionally, the authors propose a step‑aware critic (STAC) that conditions on the student’s generation step, and demonstrate that joint scaling combined with STAC consistently improves few‑step distillation for class‑conditional ImageNet generation, text‑to‑image generation with SDXL and FLUX.2, and text‑to‑video generation with Wan2.1.